YAMNet Detects Only 46% of Sudden Sounds in 48-Stream Real-Time Test

A developer tested whether Google's YAMNet audio classification model could detect unexpected sounds in a live stream without being told what to listen for. The experiment evaluated five detection methods — including spectral flux, score deltas, embedding deltas, and k-nearest-neighbour comparisons — across 48 audio streams built from ESC-50 dataset samples. The best-performing configuration identified 22 of 48 sudden sound events, achieving a recall rate of just 45.8%. While the pipeline processed audio comfortably faster than real time, the results showed that simple distance measures over YAMNet embeddings are not reliable enough for unseen sound detection. The full code and results have been published in the kiarina/labs repository on GitHub.